What problem does the Multi-Query RAG System solve?

Basic RAG implementations struggle with complex, multi-part questions. Retrieved context is often irrelevant or low quality. Poor filtering leads to hallucinations. Architecture becomes hard to maintain as the system scales.

How does the Multi-Query RAG System work?

The system receives a complex question. A decomposer agent (GPT-5-mini with memory) breaks it into 1–5 targeted sub-queries. Each sub-query retrieves relevant chunks from the Supabase vector store. Relevance filtering keeps only chunks scoring above 0.4. All surviving chunks are aggregated into a single context. A dedicated 'Think' reasoning step synthesises the aggregated context. The final answer is generated, grounded strictly in the knowledge base.

Which tools and integrations does it use?

It is built on n8n and integrates n8n, Supabase Vector Store, OpenAI GPT-5-mini, NVIDIA Embeddings, JavaScript.

What results does it deliver?

90% improvement in answer quality versus basic RAG, validated by the client. 60% reduction in hallucinations. Faster, more accurate responses through targeted sub-queries. Scales to knowledge bases with millions of vectors. Reusable template adaptable to courses, company docs, support KBs, and research tools.

Multi-Query RAG on Supabase pgvector with n8n & OpenAI

n8n RAG agent on Supabase pgvector and OpenAI: splits questions into 1-5 sub-queries, keeps chunks above 0.4 similarity and answers only from your docs.